The Video That Got Thinking Almost Right
There is a video doing the rounds called Bodybuilding for the Mind, and it makes one of the cleanest arguments I have read about what AI is doing to how we think. I want to take it seriously before I take issue with it, because the argument is strong, and the exact place it stops is the place the real work starts.
The argument runs like this. For most of history, if a task needed reasoning, a human had to supply the reasoning. That is ending. AI is making thinking optional the way machines made physical labour optional. And when a capability stops being necessary for survival, something predictable happens: it turns into a discipline. Machines took the physical work out of daily life, so nobody had to be strong to eat — and the modern gym was invented so people could voluntarily impose a load on themselves. Once the load was voluntary it could be optimised, the development of muscle became the point, and the physiques that came out of it were unlike anything that had existed before.
The video predicts the same sequence for the mind. Thinking becomes voluntary, so it becomes a discipline; AI becomes the equipment; and we get a leap in reasoning the way we got a leap in physiques. The proof it offers is chess. Once engines could calculate better than any human, the engines stopped being rivals and became training equipment — and human chess got stronger, not weaker. The warning is just as blunt. Most people will let their thinking decay, an obesity epidemic for the brain, while the few who train pull away.
I think the diagnosis is right. I think the warning is right. The model underneath it is a training model, and training is not what thinking is.
Training Assumes You Already Know the Movement
Here is the sentence that carries the whole trouble: it is the struggle that is the training signal.
That is true for a muscle. It is only half true for a mind, and the missing half is the half that matters.
Think about what a rep in the gym actually is. A bench press is a fixed movement with a known form. The skill already exists, and a coach can show you the bar path, the grip, the tempo. Your job over months is to repeat that exact form under more load, and the load is the only variable. Training, in that sense, assumes the movement is known and that all that is left is intensity.
Now ask what a rep of thinking is. There is no fixed form. When you face a genuinely new problem, nobody can show you the bar path, because the movement has not been invented yet. The skill is not known in advance; it has to be found. And when you find it, the finding changes what you can do the next time. That is not training. That is learning about learning.
This is the distinction I keep returning to. Training assumes we know the skill and need the load. Learning assumes we do not yet know the skill, so the first job is to assess our assumptions and revise how we are going about it. You can be the strongest person in the gym and still be doing the wrong movement, and no amount of extra weight will tell you, because weight cannot audit form. Only a mind can.
The video's own example gives the game away. Chess engines did not add load to a movement humans had already mastered. They forced humans to rebuild their model of the game, away from pure calculation and toward pattern, structure and intuition. The engines changed how players learned to read a position. That is modelling, not weightlifting. The analogy quietly switches from training to meta-learning at the exact moment it has to explain the leap.
Bateson Named This in 1968 — Learning I, II and III
The missing layer has a name, and it is older than NLP. In a paper first delivered in 1968 and collected in Steps to an Ecology of Mind in 1972, the anthropologist Gregory Bateson set out a hierarchy of learning that explains exactly what the gym model leaves out.
Learning I is change within a set of alternatives. You learn the correct response in a known situation. Repeat, correct, repeat. This is the level the video is describing, and it is real — it is where skill gets competent.
Learning II, which Bateson also called deutero-learning, is change in the set of alternatives itself. It is learning to learn: picking up the pattern of what kind of situation this is, what counts as a good move here, how learning works in this context. Nobody teaches it directly. It is absorbed from the shape of the situation. This is the level the chess example is really about, and the level at which a person stops being trained and starts being self-directed.
Learning III is the rare one. It is a change in the process of Learning II — a restructuring of the whole system that produced your habits of learning, which Bateson described as a profound reorganisation of character. It happens seldom, and it is uncomfortable, because it means holding contradictions in your own way of working rather than resolving them by picking a side.
NLP comes directly out of this lineage. Bateson is the man who introduced Richard Bandler and John Grinder to Milton Erickson and sat in on the early modelling sessions; the frames, the levels and the interest in paradox all trace back to him. So when someone tells me thinking is going to become a gym, I hear them describing Learning I with more weight, and I want to ask what happened to Learning II and III — the levels where the weight is not the point.
Pillar One — Holding Paradoxes Without Collapsing Into One Answer
If thinking is meta-learning, what is it that actually gets trained? I think there are three capacities, and none of them is load. The first is the ability to hold more than one contradictory approach at once without forcing an early answer.
Most of what looks like thinking is frame-selection: pick a way of looking at the problem, then optimise inside it. That is genuinely useful, and it is what most training rewards. But the higher move is to run two incompatible frames side by side and let them argue. In 1971 the psychiatrist Albert Rothenberg gave this a name — Janusian thinking, after the Roman god with two faces — defining it as the capacity to conceive and use two or more opposite or contradictory ideas simultaneously. He found it in the working notes of people who produced genuinely original work, not in the ones who merely executed well.
Half a century later this became measurable in business. Ella Miron-Spektor and colleagues, writing in the Academy of Management Journal in 2018, studied what they called a paradox mindset: the extent to which a person is energised by tensions rather than flattened by them. People who scored high on it were better able to use competing demands — efficiency against creativity, control against flexibility — to improve their actual work, rather than treating the tension as a problem to be eliminated. The finding is specific and useful: the tension itself is the productive site, and the instinct to resolve it too fast is what costs you.
This is why I do not buy the pure-load version of mind training. Loading a single frame makes you faster inside it. Holding two frames that cannot both be true at the same time, and refusing to collapse them until the problem forces the choice, is a different skill entirely, and it is the one that produces the leap the video is promising.
Pillar Two — Modelling, or Borrowing a Mind to Widen Your Own
The second capacity is modelling: importing another person's way of thinking so that your own range of possible moves expands.
The evidence that we learn by watching is old and strong. Albert Bandura's 1965 experiments on vicarious reinforcement showed that children could acquire a behaviour not by being rewarded themselves, but by watching someone else be rewarded — the observer updates without acting. This is the mechanism behind the phrase vicarious reinforcement, and it is one of the most reproduced findings in social psychology. We are built to learn from models, and we do it without having to run the experiment ourselves.
NLP took that principle and turned it into a deliberate practice. Bandler and Grinder did not invent modelling from theory; they did it. They sat with the transcripts of Fritz Perls and Virginia Satir and the recordings of Milton Erickson, and they reverse-engineered the structure of what those people did, so that a learner could borrow the structure rather than spend a lifetime reinventing it. The Meta Model came out of Perls; the Milton Model came out of Erickson. Whatever you think of the packaging, the underlying move is the same one Bandura documented: a mind learns from a mind.
This matters because it corrects the pure-repetition story in a measurable way. K. Anders Ericsson's deliberate-practice framework is the strongest case that practice, not talent, makes experts — but the meta-analysis by Brooke Macnamara and colleagues in 2014 found that deliberate practice explained only about 26 percent of performance variance in games, 21 percent in music, 18 percent in sports, 4 percent in education, and less than 1 percent in professions. I want to be careful with that number: it does not mean practice does not matter, and Ericsson argued the estimate was too low. It means that repetition inside one frame is not the whole engine. The rest of the engine is the quality of the models you have available, and your ability to stretch your thinking toward them — which is exactly what produces the instinctive, intuitive flexibility that makes someone look effortlessly fast.
Pillar Three — The Unconscious Update, or Learning Without Deciding
The third capacity is the one that runs while you are not looking: the accumulation of structure from many sources without a conscious decision at each step.
That this happens at all is well established, and it was not obvious before it was shown. In 1967 Arthur Reber exposed people to strings of letters generated by a hidden set of rules, asked them only to memorise the strings, and then tested them on new strings. They could reliably tell which new strings obeyed the hidden rules and which did not — well above chance — while being unable to say what the rules were. They had learned the grammar of a system without ever being able to state it. This is implicit learning, and the artificial-grammar paradigm Reber built is still one of the standard tools of cognitive psychology.
The unconscious update does not stop at the end of the session, either. In 2004, Ullrich Wagner and colleagues published a study in Nature titled Sleep Inspires Insight. People worked a task that could be solved slowly by grinding, or suddenly by noticing a hidden rule. After the training, some slept and some stayed awake. More than twice as many of the sleepers found the hidden rule. The sleeping brain restructures what you learned and hands you back an insight you did not consciously assemble. That is the unconscious update doing its work, and it is why the deepest learning so often arrives after you stop pushing.
Two of the mechanisms people reach for here deserve an honest note. Motor imagery — mentally rehearsing a movement — has real support: the PETTLEP framework set out by Holmes and Collins in 2001, and a body of sport research that followed, found that imagery-based rehearsal improves skilled performance, in part because imagining a movement engages some of the same systems as performing it. Visualisation, in other words, is not mysticism; it has a measured effect. Hypnosis is where I would be more careful. Its effects are real in specific, narrow domains, and it has a documented history in Erickson's work, but it is not the broad engine its reputation suggests, and I will not sell it to you as one.
Put the three together and you get the actual picture. Learning I is the load. Learning II, the level the video did not name, is where the real gain lives: holding contradictions, borrowing structures from other minds, and letting the unconscious consolidate what you never sat down to consciously decide.